For years, AI video meant silent clips and a separate trip through a voice or sound tool. MiniMax just shipped a model that treats audio as part of the picture — and posted the weights for part of it.
In 2026, MiniMax open-sourced H3, its next-generation general-purpose video model. The interesting claim is not the resolution. It is that dialogue, sound effects, and room tone come out of the same generation as the frames, timed to what happens on screen.
One pass, picture and soundtrack together
H3 is an omni-modal system: it reads text, images, video, and audio together and writes video with native 32 kHz stereo sound. There is no lip-sync pass to line up afterward and no separate music step.
Outputs run 4 to 15 seconds at 24 FPS. The default short edge is 768 pixels, with 2K reached through a second stage called H3-Regenerate-2K.
What "omni-reference" really lets you do
The Ref2VA variant is where creators will spend time. A single generation can take:
- up to 9 reference images,
- up to 3 reference video clips (each 2–15 seconds),
- up to 3 reference audio clips.
H3 carries character identity, motion, and voice across the shot. Tell it, in words, to borrow the camera move from one clip, the singer from an image, and the vocals from an audio file, and it assembles the relationship itself. That is the difference between a model that edits and a model that directs.
Stable language support covers 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.
The price the official blog quietly dropped
MiniMax's own launch post states the cost clearly: at 2K, H3's per-second price is less than a third of mainstream models, and at 768p it is less than half the price of mainstream 720p offerings. For teams producing commercial short dramas (短剧), ads, or product videos, that ratio is the part that changes budgets — not the benchmark.
The license is open — with strings
H3-Base ships as open weights on Hugging Face under the MiniMax H3 Community License. That permits building and customizing, but read it before commercial use; it is not a blanket open-source grant.
Two pieces of the full system are not in the open release:
- H3-Context-IR, the preprocessing step that turns messy multimodal prompts into something H3 can use, stays hosted. MiniMax provides an API to reproduce it.
- H3-Regenerate-2K, the step that lifts 768p output to 2K, is not yet open-sourced.
So "open H3" today means you can run H3-Base (FL2VA for text/image-to-video, Ref2VA for reference workflows) locally. The complete 2K pipeline is not fully open yet.
Where it fits
Closed-source labs have dominated video generation. H3 is the clearest sign that open weights are reaching this category too. Its pitch — instruction following, accurate text and brand rendering, and video-to-video motion transfer — targets advertising, e-commerce, product design, UI/UX, and gaming.
A practical note: because only H3-Base is open, the quality you get locally tracks the 768p base output unless you rebuild the 2K stage yourself. Teams that need the sharper 2K look today will lean on the hosted API or H3 Max, a post-trained high-speed variant MiniMax lists in its model catalog. That split — open base, hosted polish — is the real shape of open video in 2026.
Why this matters beyond China
Open video models change who gets to make films. When a lab in Shanghai releases weights anyone can run, a student in Lagos or a small studio in São Paulo can build on the same toolkit as a well-funded production house. The catch is quality control: open base weights still need hosted polish for broadcast-grade output, and that gap is where the next round of competition will play out.
Honest limitations
- Only H3-Base is open. The Context-IR preprocessor and the 2K upscaling stage are hosted or not yet released, so a fully local 2K workflow is not available today.
- The Community License restricts commercial use in ways a true open-source license would not. Check the current terms for your deployment.
- The "less than a third of mainstream" price claim is MiniMax's own comparison against unnamed competitors; independent per-second pricing varies by provider.
- Generation is capped at 15 seconds. Long-form narrative still needs stitching or external editing.
What readers can do now
- Grab the MiniMaxAI/MiniMax-H3 repository on Hugging Face and start with FL2VA for text-to-video.
- If you need character or product consistency across clips, use Ref2VA with reference images and video rather than prompts alone.
- For a fully local stack, plan to rebuild Context-IR yourself using MiniMax's prompting guidance — the hosted version is not in the weights.
- Compare the hosted API price per second against your current video tool before assuming self-hosting is cheaper; at this scale it often is not.
- Keep the license open in a tab during evaluation; commercial thresholds change, and the Community License is not Apache 2.0.